slub / ocrd-demo-2023-06-30

Demos for OCR-D presentation at Seminar Datenkompetenz für DH – OCR

Demos OCR-D

Demo 0 - install ocrd_all or ocrd/all

Install OCR-D locally

Read https://ocr-d.de/en/setup guide

Decide for native or Docker installation

Install (native)

git clone https://github.com/OCR-D/ocrd_all
cd ocrd_all
sudo make deps-ubuntu
# optionally:
sudo make deps-cuda
make all
. venv/bin/activate
mkdir -p ~/data
cd ~/data

Install (Docker)

docker pull ocrd/all:maximum
# or:
docker pull ocrd/all:maximum-cuda
mkdir -p ~/data
docker run -it -v ~/data:/data -v ocrd-models:/models ocrd/all bash

Download models

ocrd resmgr list-available
ocrd resmgr download ocrd-sbb-binarize default-2021-03-09
ocrd resmgr download ocrd-tesserocr-recognize deu.traineddata
ocrd resmgr download ocrd-tesserocr-recognize frk.traineddata
ocrd resmgr download ocrd-tesserocr-recognize GT4HistOCR.traineddata
ocrd resmgr download ocrd-tesserocr-recognize Fraktur.traineddata
ocrd resmgr download ocrd-tesserocr-recognize Latin.traineddata
ocrd resmgr download ocrd-kraken-recognize austriannewspapers.mlmodel

Demo 1 - create workspace, run Tesseract

Start with a bunch of images, create workspace, run Tesseract all-in-one

Browse Börsenblatt

→ http://digital.slub-dresden.de/id39946221X-18560530

Copy links for "Einzelseite als Bild herunterladen"

Download

wget https://digital.slub-dresden.de/data/kitodo/Brsfded_39946221X-18560530/Brsfded_39946221X-18560530_tif/jpegs/000000{01..16}.tif.original.jpg

Mkdir and chdir

mkdir -p demo1
mv *.jpg demo1
cd demo1

Import images into new workspace

ocrd-import -P . 

Minimalist workflow

ocrd process "tesserocr-recognize -I OCR-D-IMG -O OCR-D-OCR-TESS -P segmentation_level region -P textequiv_level word -P find_tables true -P model frak2021"
# or equivalently:
ocrd-tesserocr-recognize -I OCR-D-IMG -O OCR-D-OCR-TESS -P segmentation_level region -P textequiv_level word -P find_tables true -P model frak2021

The results are in the OCR-D-OCR-TESS file group / directory.

Inspect results with browse-ocrd and JPageViewer

Visualize results with browse-ocrd or PRImA PageViewer

Demo 2 - clone workspace, more complex workflow

→ Document from SLUB Digital Collection

Start with a METS from SLUB, run a more complex workflow on it

Browse back

Select the link below "OAI-Identifier" to retrieve the METS XML

→ https://digital.slub-dresden.de/oai/?verb=GetRecord&metadataPrefix=mets&identifier=oai:de:slub-dresden:db:id-39946221X-18560530

Clone as workspace

ocrd workspace -d demo2 clone "https://digital.slub-dresden.de/oai/?verb=GetRecord&metadataPrefix=mets&identifier=oai:de:slub-dresden:db:id-39946221X-18560530"

Inspect available file groups

ocrd workspace find -k fileGrp -k url -k mimetype

Download all images in ORIGINAL file group

ocrd workspace find --file-grp ORIGINAL --download

Run the same simple workflow as above

Since our file group is called ORIGINAL, OCR-D-IMG must be replaced with ORIGINAL here.

ocrd-tesserocr-recognize -I ORIGINAL -O TESSERACT -P segmentation_level region -P textequiv_level word -P find_tables true -P model frak2021

The results are in the TESSERACT file group / directory.

Run a (somewhat) suitable workflow

1. Download the existing annotations from ABBYY Cloud and repair them (removing Shape elements with wrong coordinates)

    ocrd workspace find -G FULLTEXT --download
    xmlstarlet ed --inplace -d //_:Shape FULLTEXT/*
    # ocrd workspace prune-files # delete all other files (not downloaded)
   

2. Convert ALTO to PAGE (adding correct image reference)

    ocrd-fileformat-transform -I FULLTEXT -O PAGE -P from-to "alto page"
    ocrd-segment-replace-page -I ORIGINAL,PAGE -O PAGE2 -P transform_coordinates false
   

3. Binarize and crop

    ocrd-sbb-binarize -I PAGE2 -O BINARIZED -P model default-2021-03-09
    ocrd-anybaseocr-crop -I BINARIZED -O CROPPED -P marginBottom 0.9 -P marginTop 0.1 -P marginRight 0.9 -P marginLeft 0.1
   

4. Overwrite line segmentation, repair region segmentation

    ocrd-cis-ocropy-segment -I CROPPED -O LINES -P level-of-operation region
    ocrd-segment-repair -I LINES -O REPAIR -P plausibilize true
   

5. Dewarp on line level

    ocrd-cis-ocropy-dewarp -I REPAIR -O DEWARPED
   

6. Run multiple OCR models and combine them

    ocrd-calamari-recognize -I DEWARPED -O OCR1 -P checkpoint_dir qurator-gt4histocr-1.0 -P textequiv_level glyph
    ocrd-tesserocr-recognize -I DEARPED -O OCR2 -P model frak2021 -P textequiv_level glyph
    ocrd-tesserocr-recognize -I DEARPED -O OCR3 -P model GT4HistOCR+Fraktur+frk+Latin+deu -P textequiv_level glyph
    ocrd-kraken-recognize -I DEARPED -O OCR4 -P model austriannewspapers.mlmodel
    ocrd-cor-asv-ann-align -I OCR1,OCR2,OCR3,OCR4 -O ALIGNED -P method combined
    ocrd-page-transform -I ALIGNED -O OCRX -P xsl page-remove-words.xsl
   

The results are in the OCRX file group / directory.

Demo 3 - run various OCR engines on GT and evaluate

Download GT, process with calamari, evaluate with dinglehopper and ocrd-cor-asv-ann-evaluate

NOTE This demo is just to show how to do the evaluation. The choice of OCR engines, evaluation processors and models is entirely arbitrary and should not be seen as recommendation.

Browse to and download from OCR-D GT Repo

• Go to https://github.com/OCR-D/gt_structure_text
• Copy link to https://github.com/OCR-D/gt_structure_text/releases/download/v1.4.8/luz_blitz_1784.ocrd.zip

wget https://github.com/OCR-D/gt_structure_text/releases/download/v1.4.8/luz_blitz_1784.ocrd.zip

Extract the OCRD-ZIP

Extract the data subdirectory of the ZIP (which contains the workspace)

unzip luz_blitz_1784.ocrd.zip 'data/*'

Run small workflows for OCR results with Tesseract and Calamari, compare output

This workflow uses ocrd-olena-binarize (with the sauvola-ms-split algorithm) to binarize the images. The images are processed by two runs with Tesseract (using Fraktur_GT4HistOCR and deu) and one run with calamari (using qurator-gt4histocr-1.0).

ocrd process -m data/mets.xml \
  "olena-binarize -I OCR-D-GT-SEG-LINE -O BIN" \
  "tesserocr-recognize -P segmentation_level word -P textequiv_level line -P find_tables true -P model frak2021 -I BIN -O TESS-GT4HIST" \
  "tesserocr-recognize -P segmentation_level word -P textequiv_level line -P find_tables true -P model deu -I BIN -O TESS-DEU" \
  "calamari-recognize -P checkpoint_dir qurator-gt4histocr-1.0 -I BIN -O CALA-GT4HIST"

This allows us to compare the files in TESS-GT4HIST, TESS-DEU and CALA-GT4HIST with each other and with the GT in OCR-D-GT-SEG-LINE.

Compare all the OCR results with the GT using ocrd-cor-asv-ann-evaluate

ocrd-cor-asv-ann-evaluate -m data/mets.xml -I OCR-D-GT-SEG-LINE,TESS-GT4HIST,TESS-DEU,CALA-GT4HIST -O EVAL-ASV -P confusion 20 -P metric Levenshtein

The results are JSON files in the EVAL-ASV filegroup with line-by-line distance measures between all the engine.

data/EVAL-ASV/EVAL-ASV.json contains the metrics (CER mean and variance) and top confusion table for the full workspace:

{
  "OCR-D-GT-SEG-LINE,TESS-GT4HIST": {
    "length": 110,
    "distance-mean": 0.0368893299998711,
    "distance-varia": 0.011348232597131346,
    "confusion": "([(20, ('\u2e17', '-')), (6, (0, '*')), (6, (0, ' ')), (5, ('\u017f', 's')), (4, (0, '-')), (4, (0, 'i')), (4, (0, 'e')), (3, ('\u2014', '-')), (3, (0, '.')), (3, (0, 'r')), (2, ('u', 'l')), (2, ('R', 'K')), (2, (0, '\u017f')), (2, (0, 'c')), (2, ('\u2e17', ' ')), (2, (' ', 0)), (1, ('\u201c', 0)), (1, ('t', 'r')), (1, ('.', '-')), (1, ('3', 0))], 4425)"
  },
  "OCR-D-GT-SEG-LINE,TESS-DEU": {
    "length": 110,
    "distance-mean": 0.17129872156596126,
    "distance-varia": 0.03381635236354291,
    "confusion": "([(100, ('f', '\u017f')), (16, (' ', 0)), (15, (',', '.')), (11, ('\u00fc', 'u\u0364')), (10, ('\u00f6', 'o\u0364')), (10, ('f', 0)), (9, ('b', 'd')), (9, ('\u00e4', 'a\u0364')), (9, ('s', '-')), (9, (0, '-')), (8, ('S', 'G')), (8, ('f', 'k')), (7, ('r', 't')), (7, (0, ' ')), (6, ('(', '\u017f')), (6, ('u', 'u\u0364')), (6, ('b', 'h')), (5, (0, '\u017f')), (5, ('M', 0)), (5, (0, 'z'))], 4510)"
  },
  "OCR-D-GT-SEG-LINE,CALA-GT4HIST": {
    "length": 110,
    "distance-mean": 0.05106674581570344,
    "distance-varia": 0.01795657646161947,
    "confusion": "([(20, ('\u2e17', '-')), (17, (0, 'e')), (12, (0, '*')), (11, (0, ' ')), (8, (0, 'l')), (7, (0, 'c')), (7, (0, 'r')), (6, ('e', 'c')), (6, (' ', 0)), (6, (0, 'd')), (5, (0, ',')), (5, (0, 'h')), (5, (0, 'i')), (5, (0, 'n')), (4, ('.', '-')), (4, ('\u017f', 's')), (4, (0, '-')), (4, (0, 'u')), (3, (0, 'S')), (3, (0, 'a'))], 4429)"
  }
}

data/EVAL-ASV/EVAL-ASV_0003.json contains the metrics for each line of page 3.

Compare Calamari output with GT using ocrd-dinglehopper

ocrd-dinglehopper -m data/mets.xml -P textequiv_level line -I OCR-D-GT-SEG-LINE,CALA-GT4HIST -O EVAL-DINGLE

The result are HTML files (Diff View) and JSON files (with CER and WER).

HTML for page 3:

JSON for page 3:

{
    "gt": "OCR-D-GT-SEG-LINE/OCR-D-GT-SEG-LINE_0003.xml",
    "ocr": "CALA-GT4HIST/CALA-GT4HIST_0003.xml",

    "cer": 0.07770472205618649,
    "wer": 0.1320754716981132,

    "n_characters": 1673,
    "n_words": 265
}

Visualize with browse-ocrd

Show diff view in browse-ocrd (https://github.com/hnesk/browse-ocrd/tree/diff-view)

browse-ocrd data/mets.xml

Demo 4 - makefiles

Recreate demo1 and demo2 via equivalent makefiles

Try to build

ocrd-make -f demo1.mk demo1
ocrd-make -f demo2.mk demo2

make[1]: Entering directory 'demo1' make[1]: 'OCR-D-OCR-TESS' is up to date. make[1]: Leaving directory 'demo1' make[1]: Entering directory 'demo2' make[1]: OCRX is up to date. make[1]: Leaving directory 'demo2'

Trigger rebuild

touch demo1/OCR-D-IMG demo2/ORIGINAL
ocrd-make -f demo1.mk demo1
ocrd-make -f demo2.mk demo2